assistant professor, smaranda cimpoeru, phd e-mail ...ecocyb.ase.ro/nr20152/05 -...
TRANSCRIPT
-
Assistant professor, Smaranda CIMPOERU, PhD
E-mail: [email protected]
The Bucharest University of Economic Studies
USING SELF-ORGANIZING MAPS FOR ASSESSING SYSTEMIC
RISK. EVIDENCES FROM THE GLOBAL ECONOMIC CRISIS
Abstract. The importance of correctly assessing systemic risk has
increased substantially after the events from 2007. A wide set of parametric and
non-parametric methods has been used to address the problem and identify the
leading risk factors for potential crisis situations. Out of these, the class of neural
networks represented by self-organizing maps has become an important technique
but only with a modest usage for the economic crisis. We apply the self-organizing
maps technique for a set of worlds’ economies, with the goal of identifying the
resemblances and differences between worlds’ economies. The model developed on
self-organizing maps ensures detection of the imbalances and vulnerabilities of
economies and find the determinant variables (early warning signals) for a
financial-economic crisis situation. The study has the advantage of including in the
analysis a pre-crisis and a post-crisis assessment, gaining much insight from the
structural changes produced in the topology of the economies.
Keywords: Self-Organizing Maps, Neural Networks, Early Warning
Systems, Systemic Risk, Global economic crisis.
JEL Classification: C45, H63, C49
1. Introduction
The globalized financial environment that we are experiencing nowadays
creates the premises for the financial instability to be transmitted over the countries
which could lead to a generalized collapse of the real economy. Apparently,
although the stress tests performed on European level gave good results, the credit
ratings for many countries in Europe worsened during 2011. Miricescu (2014)
makes an analysis of the dominant factors that impact the long-term sovereign
rating. In his paper, he highlights the government debt to GDP ratio which raised
for the EU from app. 62% in 2000 to 88% in 2014. In this context, the problem of
finding and evaluating an accurate early warning system for the European financial
system becomes a real challenge and of utmost importance. Many papers have
investigated what were the causes of the crisis. For example, Reinhart and Rogoff
(2008) find the following lead indicators of crisis: real housing and equity prices,
current account deficits, GDP growth, increases in debt.
mailto:[email protected]
-
Smaranda Cimpoeru
_________________________________________________________________ Recent financial crisis has demonstrate that the economies are vulnerable in front
of the systemic financial distress and that it very important for policy makers to
understand the sources of these vulnerabilities in order to increase the capacity for
absorbing shocks of the financial and economic system. Moreover, the financial
integration which is a very wide phenomena (Moscalu, 2014) leads to significant
linkages across markets, so that a holistic approach of the financial crisis has to be
applied.
Especially after the eruption of the global crisis in 2007, the importance of
understanding and correctly assessing risk factors and risk transmissions within
markets of the economy has risen as a particularity of assessing financial instability
situations. As a general definition (Oet et al, 2010), the early warning systems are
“data-driven approaches” with the goal of identifying variables associated with past
crises and alert policy makers of other potential future crises. Early warning
systems are based on the assumption that crisis factors can be identified before a
crisis and can be used for improving policy measures at macroeconomic level.
Objective of the paper is twofold: identification of resemblances and differences
between different economies of the world in order to detect the imbalances and
adjust the corrective policy as to take into account this disequilibrium; find the
determinant variables (early warning signals) for a financial-economic crisis
situation.
Structure of the paper is as follows. In section 2 we review the specialty literature
of using Self-Organizing maps in the context of the global financial crisis. In
section 3 we introduce the basics of the self-organizing maps (including vector
quantization concept) and in the next section we present briefly the algorithm of
the method. The second part of the paper is dedicated to the case study – we
introduce the database, the variables, the topology of the economies before the
crisis (2007) and the structural mutations that took place in the post-crisis context.
Last section draws the conclusions.
2. Using Self-Organizing Maps for assessing the global financial crisis – Literature review
The main advantage of the neural networks is the fact that they are non-parametric
models that do not require the assumptions for statistical data distribution and are
not limited by linear specifications. Considering that the indicators of a financial
crisis are non-linearly related (Fioramonti, 2008), neural networks were widely
used for evaluating the financial, debt or currency crisis. However, the focus of the
current paper is on the self-organizing maps, as a class of the neural networks
models, so we will provide the literature review of using this technique in assessing
financial stress. Despite the large number of papers that study the use of SOM in
engineering or medicine, the specialty literature is whatsoever scarce in what
concerns the applications of SOM in financial stability and economic crisis
assessment.
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
First studies that apply SOM to model currency crisis are that of Arciniegas and
Arciniegas Rueda (2009). They explore the correlation between real effects of
speculative attacks on currency and a set of macroeconomic variables. In Sarlin
and Marghescu (2011) we have the same idea, of applying the SOM for the
indicators of a currency crisis. Dattels et al. (2010) develop a Global Financial
Stability Map, by using six composite indices. However, it has the disadvantage
that the sources of individual stress are difficult to identify and, as stated by the
authors, the results are to be viewed as illustrative.
Sarlin, Peltonen (2011) develop a Self-Organizing Financial Stability Map
(SOFSM) which allows the identification of the vulnerability sources and performs
well for out-of-sample systemic financial crisis. For the topologically ordered
SOFSM, the financial stability neighborhood represents the “contagion of
instabilities through similarities in the current macro-financial conditions”. The
map is represented in the following areas of the financial stability cycle: pre-crisis,
crisis, post-crisis and tranquil state. The SOFSM developed performs better than a
logit model for classifying in sample data and for predicting the global crisis from
2007. However, the map does not show the imbalances between different
economies across the world facing the economic crises.
Itturiaga and Sanz (2013) propose a model for detecting and managing divergences
between countries in order to anticipate the danger of a financial crisis. They use
the Self-Organizing Maps model to perform a classification of the European
countries, as well as German and Spanish regions. The reason for choosing these
two countries comes from the similar territorial organization but with the
significantly different financial status. The SOM model is applied to find the extent
to which national financial instability is due to the regional macroeconomic
imbalances. Public expenditure and saving rate are found to be the most critical
variables with impact on a country’s economy.
A worth to mention adaption of the Kohonen standard SOM is the Self-Organizing
Time Map designed by Sarlin (2013a, 2013b) with the purpose of abstraction the
structure in temporal multivariate problems. When ordering ascending of time the
one-dimensional arrays, the SOTM will enable a two-dimensional representation
with the multivariate data structures on the vertical and the temporal direction on
the horizontal. The ordered SOTM can be used for projecting individual or grouped
data onto the map. In Sarlin (2013b), the SOTM is applied to financial stability
surveillance. The results show that high equity prices, current account deficits and
GDP growth are the main triggers for financial crises. The Self-Organizing Time
Maps can be used for: identifying the imbalances in indicators over time, the
structural changes in data, the specific location of univariate and multivariate
changes across the data.
As also it was mentioned in Sarlin, Peltonen (2011), the SOM applied for assessing
financial and economic crisis “enables disentangling the specific threats, risks and
-
Smaranda Cimpoeru
_________________________________________________________________ triggers, and should be treated as a starting point rather than an ending point for
financial stability analysis”.
3. Self-Organizing Maps – Essentials
Neural networks can be classified in two major classes: supervised and
unsupervised networks. For the supervised ones, a target vector is presented to the
network so that it adjusts the results to the expected output. The unsupervised
networks are assimilated to the exploratory analysis and clustering methods. The
Self-Organizing Map is an unsupervised competitive type of network.
The first scientists to delimitate the notions of brain maps are Mountcastle (1957)
and Hubel and Wiesel (1962). They found that certain neural cells in the brain react
to specific sensorial stimulations. Moreover, the designated cells are grouped in
local assemblies and their location is assigned with the response to a certain
stimulus. This is the way the brain maps, which are nothing less than systems of
cells, have been discovered and defined.
Later on, Merzenich et al. (1983) has reported that the brain maps depend strongly
on sensorial experiences. This idea has developed into the competitively learning
neural networks. This means that in a sequel of cells, the process of cells’
adaptation to the input signal makes them dependent on the specific input
characteristics.
However, the brain maps models which were inspired by biology could not be
applied to data analysis, and this was mainly due to the fact that the resulting maps
were partitioned, meaning that they were made of several patches, between which
the ordering was random and discontinue, so no global order existed over the entire
map.
That is why, in the neural models used in data analysis, controlling the nodes
activities through the neural connections is not enough. There is a need for an extra
control, using factors to intermediate the information without mediating the
activities. This is where the vector quantization comes into place.
The idea of vector quantization (VQ) dates back to 1850 (Dirichlet) and 1907
(Voronoi tessellation in spaces of arbitrary dimensionality). This technology used
in digital signal processing, partitions the vector-values input data into a finite
number of contiguous regions, where each region is represented by the single
model vector or the codebook vector. The VQ is usually illustrated with the
Euclidean distance. For instance, if we consider the input data formed of n-
dimensional Euclidean vectors, denoted by Y and the model vectors denoted by 𝑀𝑖 . We denote with 𝑀𝑤 the winner model vector, the vector with the smallest Euclidean distance from the input vector, Y. Mathematically, we can write:
𝑤 = 𝑎𝑟𝑔 min𝑖
{‖𝑌 − 𝑀𝑖‖} (1)
If we further denote with f(Y) the probability density of Y, the mean quantization
error E is then defined as:
𝐸 = ∫ ‖𝑌 − 𝑀𝑤‖2𝑓(𝑌)𝑑𝑉
𝑉 (2)
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
Where dV is a volume differential on the data space V. The objective function, E, is
an energy function which could be minimized by the gradient descent procedure.
(Kohonen, 2013).
The above outlined VQ technique is also called the “k-means clustering” and the
self-organizing map can be viewed as a generalization of the k-means clustering
algorithm. The self-organizing map is first introduced by Kohonen at the beginning
of the 80s. The technique resembles the VQ, but the vector models are spatially and
globally ordered. In Figure 1, we represent a self-organizing map – for an input
vector Y, the feature map finds the winning node in the finite output space. The
associated weight vector give the coordinates of the node from the input space.
As per Kohonen (2013), the input data, Y, is mapped to a set of models (𝑀𝑖) where 𝑀𝑤is the best match for Y. All models that are in the close surrounding of the winner model are better match with Y than the others. The figure illustrates the
basis of the Self-Organizing Maps (SOM) algorithms. Similar models will be
assigned with nodes that are closer in the grid (smaller Euclidean distance in the
VQ), while less similar models will localized further away. The essentials of the
SOM, as stated by Kohonen (2013) is: “Every input data item shall select the
model that matches best with the input item, and this model, as well as a subset of
its spatial neighbors in the grid, shall be modified for better matching”. The
mentioned modification is associated with the winner model. However, due to the
fact thatit is not a single vector that changes, but an entire family of neighbors
vectors, this implies a local ordering of the models in the neighborhood. This local
ordering will propagate across the grid.
Figure 1 – Feature Map – Self Organizing Map representation
Continuous
Input Space
Finite Output
Space
Feature Map
Y Model 𝑀𝑤 with the winning
neuron and the
neighborhood
-
Smaranda Cimpoeru
_________________________________________________________________ Two algorithms can be used for producing an ordered set of models in the map.
The first one assumes a stepwise procedure, meaning that the input data are
presented to the network one at a time, for as many iterations as necessary to reach
a state of equilibrium. In the second type of algorithm, all input data are presented
to the algorithm as one batch and all the models are modified in a single operation.
The batch process repeats a certain number of times until the exact stabilization of
the models. In the second type of algorithm, the state of equilibrium is attained
faster than in the first one.
4. The Algorithm used for training the Self-Organizing Map
The basic Kohonen network has a layer of input nodes and only one layer of output
neurons. The neurons from the first layer form a discrete topological mapping of an
input space, 𝑌𝜖R𝑛. The algorithm is based on minimizing the distances between the nodes, and as mentioned before on the Vector Quantization idea.
The initial step (step zero) of the training algorithm consists in initializing all
weights of the network {𝑤1, 𝑤2, … , 𝑤𝑀}with small random values. We mention that: 𝑤𝑖 is a weight vector associated with the neuron “i”, having the same dimension, n, as the input vectors; M is the total number of neurons in the input
space and suppose that 𝑙𝑖 is the location of vector “i”on the grid. For the first phase of the algorithm, in the first step an input training vector, Y(t) is
chosen from the input space and presented to the grid. The second step of the
algorithm consists in examining every node of the network and finding the winning
neuron, that is, the neuron having the weight vector closest to the input vector, in
terms of distances (for example, in eq. 3 the Euclidean distance is used):
min 𝑑𝑗(𝑦) = √∑ (𝑦𝑖 − 𝑤𝑗𝑖)2𝑛
𝑖=1 (3)
𝑣(𝑡) = arg min𝑘∈Ω
‖𝑌(𝑡) − 𝑤𝑘(𝑡)‖ (4)
whereΩ is a set of neuron indexes. In the second phase of the algorithm, the weights of the winning neuron and its
neighbors are updated as stated in equations 5 and 6.
𝑤𝑘(𝑡 + 1) = 𝑤𝑘(𝑡) + 𝐿(𝑡)𝑛(𝑣, 𝑘, 𝑡)[𝑌(𝑡) − 𝑤𝑣(𝑡)] (5)
or otherwise stated:
∆𝑤𝑘(𝑡) = 𝐿(𝑡)𝑛(𝑣, 𝑘, 𝑡)[𝑌(𝑡) − 𝑤𝑣(𝑡)] (6)
Where:
𝐿(𝑡)is the learning rate of the network. These coefficients are scalar-valued that decrease monotonically and satisfy the following properties (7):
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
0 < 𝐿(𝑡) < 1; lim𝑡→∞
∑ 𝐿(𝑡) → ∞; lim𝑡→∞
∑ 𝐿2(𝑡) < ∞ (7)
𝑛(𝑣, 𝑘, 𝑡) is the neighborhood function, which in practice has the form of a Gaussian function, as in (8):
𝑛(𝑣, 𝑘, 𝑡) = exp [−‖𝑙𝑣−𝑙𝑘‖
2
2𝜎(𝑡)2] (8)
Where 𝑙𝑖 is the location vector of neuron “i” and 𝜎 represents the range or the radius of the neighborhood, which decrease monotonically with time. The Gaussian
function is introduced in the adjusting equation inorder to give lower importance to
the neurons of the neighborhood that are situated further away from the main node
found as the BMU (Best Matching Unit) of the input neuron.
The algorithm is than reiterated from the first step (choosing the input neuron) until
the map converges. The algorithm was presented as stated by Kohonen (2013) and
Yin (2008).
After the convergence of the SOM algorithm, the feature map has important
statistical properties (these are also mentioned in some Lecture Notes from
J.Bullinaria, 2004). First of all, the feature map which consists basically of a set of
weights in the output space offers a good approximation of the input space. This is
exactly the basis of the Vector Quantization theory that we explained in the
previous section and which is the foundation of the dimensionality reduction
process.
Secondly, the feature map is topologically ordered, meaning that the space of a
neuron in the output layer corresponds to a certain partition (feature) of the input
neurons. This is an immediate consequence of the weight update equation (eq. 4),
which is applied to all the nodes of the neighborhood and not only to the winning
neuron. That is why, the map is considered an “elastic” net – as the neuron in one
neighborhood are connected through the correspondents in the input space, than the
network will offer an image that is linked to the topological ordering of each stage
of the network training.
The third property of the feature map refers to the variation in the input
distribution. The regions where training vectors are drawn with high probability of
occurrence will be mapped into wider partitions of the output layer and with a
better resolution. Property three of the feature map will be illustrated in the case
study with the outliers distribution on the network.
The fourth property of the feature map states that the Self-Organizing Maps are
able to select the best features irrespective of the distribution in the input space,
that is they perform very well even for non-linear data. This is a very important
highlight, especially compared to other dimensionality reduction techniques, like
Principal Component Analysis which can be applied only if the data is linear.
-
Smaranda Cimpoeru
_________________________________________________________________ Otherwise said, the SOM can be viewed as an non-linear generalization of the
Principal Component Analysis, as it offers a solution for finding principal surfaces
or curves.
5. Case study
In the case study, we propose applying the Self-Organizing Map model to a set of
world economies, for two key periods: 2007 – the year before the eruption of the
crisis and 2010, the aftermath of the crisis. The inputs of the network are a set of
macroeconomic variables registered in the two periods. The results of the map are
numerous. First of all we obtain the topology of the economies before the eruption
of the crisis and the structural changes that appeared after the crisis,that is how the
world map changed from an economic point of view. Secondly, we analyze the
distribution of the macroeconomic variables across the countries for the two
periods and find the early warning signals of a crisis.
5.1 Data Base
The Data Base constructed is formed of 15 variables, which are detailed in Table 1.
Data sources include: World Bank, CIA World Fact Book, Eurostat. Decision upon
variables included in analysis follows the specialty literature, like Sarlin and
Peltonen (2011), Iturriaga and Sanz (2013).
Table 1 – Macroeconomic variables included in analysis (source of metadata:
World Bank)
V1 Agriculture, value added (% of GDP)
Agriculture corresponds to ISIC divisions 1-5 and includes forestry,
hunting, and fishing, as well as cultivation of crops and livestock
production.
V2 Cash surplus/deficit (% of GDP)
Cash surplus or deficit is revenue (including grants) minus expense,
minus net acquisition of nonfinancial assets. This cash surplus or deficit
is closest to the earlier overall budget balance.
V3 Domestic credit provided by financial sector (% of GDP)
Domestic credit provided by the financial sector includes all credit to
various sectors on a gross basis, with the exception of credit to the
central government, which is net.
V4 GDP per capita (current US$)
GDP per capita is gross domestic product divided by midyear population.
V5 GDP growth (annual %)
Annual percentage growth rate of GDP at market prices based on
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
constant local currency.
V6 Central government debt, total (% of GDP)
Debt is the entire stock of direct government fixed-term contractual
obligations to others outstanding on a particular date.
V7 Gross savings (% of GDP)
Gross savings are calculated as gross national income less total
consumption, plus net transfers.
V8 Industry, value added (% of GDP)
Industry comprises value added in mining, manufacturing (also reported
as a separate subgroup), construction, electricity, water, and gas. Value
added is the net output of a sector after adding up all outputs and
subtracting intermediate inputs.
V9 Inflation, consumer prices (annual %)
Inflation as measured by the consumer price index reflects the annual
percentage change in the cost to the average consumer of acquiring a
basket of goods and services.
V10 Interest rate spread (lending rate minus deposit rate, %)
Interest rate spread is the interest rate charged by banks on loans to
private sector customers minus the interest rate paid by commercial or
similar banks for demand, time, or savings deposits.
V11 Money and quasi money growth (annual %)
Average annual growth rate in money and quasi money. Money and
quasi money is frequently called M2.
V12 Market capitalization of listed companies (% of GDP)
Market capitalization (also known as market value) is the share price
times the number of shares outstanding.
V13 Bank nonperforming loans to total gross loans (%)
Bank nonperforming loans to total gross loans are the value of
nonperforming loans divided by the total value of the loan portfolio
(including nonperforming loans before the deduction of specific loan-
loss provisions).
V14 Stocks traded, total value (% of GDP)
Stocks traded refers to the total value of shares traded during the period.
This indicator complements the market capitalization ratio by showing
whether market size is matched by trading.
V15 Unemployment, total (% of total labor force) (modeled ILO estimate)
Unemployment refers to the share of the labor force that is without work
but available for and seeking employment.
-
Smaranda Cimpoeru
_________________________________________________________________ The variables from Table 1 are recorded for a sample of 80 countries1, chosen
based on the percentage in global GDP and availability of the data .Although we
identified a set of outliers in the data, we decided to keep the respective economies
in the analysis considering the importance of the respective economies for the
study. We mention that Macedonia, Bosnia and Herzegovina and Armenia are
outliers (upper limit) for unemployment, with the highest values from the series.
The economy of Qatar records an extremely high value for the Value added in
industry, while Norway is situated at the upper limit of the Credit surplus and
Canada at the lower limit of Money growth. On the other hand, Nigeria is situated
at the other extreme, with a very high value for the Money growth and Ukraine at
the upper limit of the rate for Non-performing loans. We will observe that these
outliers will be counted as such also in the construction of the map.
5.2 State of the economies in 2007 determined by the SOM model
Due to lack of data for the entire sample, the variables Government Debt (V6) and
the Interest rate spread (V10) were removed from the list of inputs. We use the
variables registered at 2007 and after applying the algorithm, we obtain the map in
Figure 2. The80 countries from the sample are grouped on 9 regions, three
dominant ones comprising 62 countries out of the total.
We will start by analyzing the first group of countries (center, blue in Figure 3)
which is also the most numerous. The 26 countries can be classified geographically
as follows:
o Latin America: Chile, Bolivia, Peru, Argentina, Mexico, Colombia, Uruguay, Guatemala, El Salvador, Brazil.
1Argentina, Armenia, Australia, Austria, Belarus, Belgium, Bolivia, Bosnia and
Herzegovina, Brazil, Bulgaria, Canada, Chile, China, Colombia, Cote d’Ivoire, Croatia,
Cyprus, Czech Republic, Denmark, Dominican Republic, Egypt, El Salvador, Estonia,
Finland, France, Georgia, Germany, Greece, Guatemala, Hungary, Iceland, India,
Indonesia, Iran, Ireland, Italy, Japan, Jordan, Korea Rep., Latvia, Lithuania, Macedonia
FYR, Malaysia, Malta, Mexico, Moldova, Morocco, Netherlands, New Zealand, Nigeria,
Norway, Oman, Pakistan, Paraguay, Peru, Philippines, Poland, Portugal, Qatar, Romania,
Russian Federation, Saudi Arabia, Serbia, Singapore, Slovak Republic, Slovenia, South
Africa, Spain Sri Lanka, Sweden, Switzerland, Thailand, Trinidad Tobago, Tunisia,
Turkey, Ukraine, United Kingdom, United States, Uruguay, Zambia.
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
o Europe: Slovenia, Romania, Czech Rep., Poland, Slovak Rep., Estonia, Lithuania, Portugal, Latvia, Bulgaria, Croatia, Greece,
Hungary, Malta.
o Asia: Indonesia, Turkey. We might say that, from a geographic point of view, there is a predominance from
Central and East European countries and Latin American ones. We will now
analyze the characteristics of the first group of countries. The variables for which
the mean in the group is significantly different (lower in this case) than that of the
entire sample are: market capitalization (V12), Stocks traded (V14), Gross savings
(V7) and GDP/capita (V4). Considering the variables that individualize this group
of countries, we can outline the following traits: capital market insufficiently
developed (stocks traded, market capitalization), low financial power of the
population (GDP/capita and gross saving). After exposing the characteristics for all
the group of countries, we will analyze the position of the neighborhoods on the
map and what important conclusions can be drawn.
Figure 2 – Self-Organizing Map of the economies in 2007
We continue our exposure for the so called clusters with the group on the left (the
red one). The 21 countries included in this group are the following (based on the
geographic criterion):
-
Smaranda Cimpoeru
_________________________________________________________________ o Europe: Island, Switzerland, Sweden, Finland, UK, Spain,
Denmark, Ireland, Netherlands, Germany, Austria, Belgium, Italy,
Cyprus, France.
o Asia & Australia: Korea, Japan, Australia, New Zeeland. o North America: USA, Canada.
In Figure 3 we have the characteristics of the group. We find that the designated
countries register a value higher than the rest of the economies for the following
variables: GDP/capita (V4), Domestic Credit (V3), Stocks traded (V14), and
Market capitalization (V12). While the variables that register a lower average than
that of the group are: Agriculture value added (V1), Inflation (V9), GDP growth
(V5), Money growth (V11) and Industry value added (V8). We could say that this
is the group of developed economies, with a mature financial market, however
threatened by a stagnation of the economy (GDP growth, Money growth on a
negative path).
We note that in the process of training the map, we used the standardized GDP and
in the figures below, the initial values are presented. The software used was
ViscoverySOMine, with the courtesy of the producers, in the purpose of academic
research.
Figure 3 – Characteristics of the second group of countries
In the left part of the map, we also find two smaller subsets: Jordan and South
Africa (left, lower corner) and Norway plus Singapore (left, upper cornet). Norway
and Singapore are characterized by the highest values for Cash surplus (V2), Gross
Savings (V7) and GDP per capita (V4), while Jordan and South Africa have the
greatest market capitalization. Based on these extremes, it becomes easier to
compare the countries from the red cluster considering their neighbors. We will
return to this issue at the end of the section.
The third group of countries includes the following 15 economies (the yellow
group on the right), classified on the geographical criterion:
o Europe: Serbia, Georgia, Moldova. o Asia: Iran, Belarus, Russia, Pakistan, Sri Lanka. o Africa: Egypt, Tunisia, Zambia, Cote d’Ivoire, Nigeria. o Latin America: Rep. Dominican, Paraguay.
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
In Figure 4, we find that the variables with a higher average than that of the entire
sample are: value added in the agriculture (V1), Inflation (V9), Money growth
(V11), while the variables with a lower average than that of the sample are:
Domestic Credit (V3), GDP per capita (V4) and Stocks traded (V14). This could
mean that this group of countries are characterized by a moderate economic
development and a low development of the financial market (Stocks traded,
domestic credit).
Figure 4 – Characteristics of the third group of countries
Still on the right side of the map we find Bosnia and Herzegovina, Macedonia and
Armenia, grouped based on the very high levels for the unemployment in these
economies. We also mention in this side of the map Ukraine, with an extreme high
value for the level of non-performing loans.
The last group of countries (upper side of the map, green) includes the nine
following countries:
- Asia: China, Philippines, Saudi Arabia, Oman, Malaysia, India, Thailand. - Africa: Trinidad and Tobago, Morocco.
This group is individualized by higher values for Gross Savings and Value Added
in Industry, thus could be assimilated to the strongly industrialized countries. In
Figure 5 we have the topology of the 13 variables included in the analysis on the
map of countries – these are the so called U-matrixes.
Figure 5 – Distribution of variables (U-Matrix) at 2007
-
Smaranda Cimpoeru
_________________________________________________________________
Based on Figure 5 and on the characteristics identified above, we can make a
comparative analysis of the economies before the crisis (2007). Starting from the
left side of the map, we can say that on this side we have concentrated the
developed economies, as shown by the distribution of all macroeconomic variables
(Figure 5). We can say that in 2007, the best situated economies were that of
Sweden, Norway, Finland, USA, Great Britain, Netherlands, Canada. For Japan
and Island, we notice the “red” zones on the Domestic credit, signaling very high
levels of this variable. Ireland, Spain and Italy, on the other hand are closer to the
“blue” group of countries (Cluster 1), meaning that they have characteristics
similar to the countries situated in the center of the map. We observe that in the
center of the map we have mostly the economies from South America and from
Central Europe (Poland, Czech, Croatia, Slovenia) and a key characteristic, as seen
from Figure 5, is the increased current account deficit at macroeconomic level. A
subgroup of countries from the first cluster are more oriented to the right side of
the map, that is to the countries with the lowest economic development from the
sample. In this category we have the Baltic countries, Romania, Argentina,
Colombia, Bulgaria, Turkey. These are the countries where we also register a low
level of the development for the financial and moreover for the capital market. We
also highlight the large non-performing loans rate for Egypt, Tunisia and Ukraine,
while the top countries as for value added industry are situated at the top of the
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
map (China, Malaysia, Oman, Saudi Arabia). This is the “picture” of the worlds’
economies in the year preceding the global financial and economic crisis. In the
next section we will analyze the mutations produced among the economies in the
aftermath of the crisis.
5.3 Structural changes in the map at 2010
In this section of the paper we will analyze the self-organizing map in 2010, based
on the same inputs used for 2007 (sample was reduced with four countries due to
lack of data). In Figure 6 we have the topology of the economies on four distinct
groups.
Figure 6 – Self-Organizing Map of the economies in 2010
The first group includes the following countries (the “blue” group in Figure 6):
- Mexico, Colombia, El Salvador. - Poland, Slovenia, Czech Republic, Slovakia, Finland, Belgium, Austria,
Germany, France, Portugal, Ireland, Greece, Island, Malta, Italy, Romania,
Bulgaria, Hungary, Estonia, Croatia, Lithuania, Latvia, Macedonia, Serbia,
Bosnia and Herzegovina, Cyprus
- New Zeeland, Japan, Morocco, Jordan, Tunisia
-
Smaranda Cimpoeru
_________________________________________________________________ The main characteristic for this group of countries is the low level for the GDP
growth. The “red” group of countries is individualized by high levels of inflation,
money growth, agriculture value added and low levels of GDP/capita and Domestic
Credit. Based on this descriptions, we can conclude that these are the countries
which were the most affected by the Global crisis. Countries included in the “red”
group are:
- Bolivia, Peru, Uruguay, Argentina, Paraguay, Brazil. - Indonesia, India, Sri Lanka, Pakistan, Egypt, Turkey, Georgia, Russia,
Armenia, Ukraine, Moldova.
- Nigeria, Zambia, Cote d’Ivoire. On the right side of the map we find two sets of countries. Namely, the “yellow”
and the “green” group. The “yellow” countries have high values for the value
added in industry and for Savings. These countries are the following:
- Qatar, Oman, Saudi Arabia, Thailand, Philippine, Singapore, Korea, Malaysia.
- Norway, Chile, Trinidad Tobago. The last group of countries registers high values for the market capitalization,
Stocks traded, Domestic credit and for the GDP/capita. The countries in this group
are:
- Sweden, Denmark, Great Britain, Switzerland, Netherlands, Spain. - Canada, USA, Australia, South Africa.
In Figure 7 below, we can observe the distribution of the variables in 2010 for the
entire sample.
Figure 7 – Distribution of variables (U-Matrix) at 2010
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
6. Conclusions
In the present paper we have focused on two main objectives – observe the
way the major events in 2007 affected a whatsoever “stable” map of the economies
and focus on the key drivers that contributed to the overall performance of the
economies during the crisis period, that is delimiting some “early warning signs”
of crisis. We find that developed economies in the Western Europe have been as
vulnerable as the emerging ones, in terms of macroeconomic indicators
performance. The economies which that could be considered the less affected by
the global crisis are the Asian markets and some Northern Europe economies. This
result can be put on the account of the confidence in the financial markets of the
Asian economies, which confirms the importance of the financial stability in the
global economic environment.
The paper enriches the specialty literature of using self-organizing maps for
characterizing crisis situations, by adding significant insight into the changes that
took place on the “map” of economies after the global crisis from 2007, by
identifying main groups of countries, their particularities and the distribution on the
considered macroeconomic variables.
-
Smaranda Cimpoeru
_________________________________________________________________
Acknowledgments
This work was cofinanced from the European Social Fund through Sectoral
Operational Program Human Resources Development 2007-2013, project
number POSDRU/159/1.5/S/134197 „Performance and excellence in doctoral
and postdoctoral research in Romanian economics science domain”.
REFERENCES
[1] Arciniegas Rueda, I.E., Arciniegas, F.(2009), SOM-based Data Analysis of Speculative Attacks’ real Effects.Intelligent Data Analysis, 13(2), pp
261–300;
[2] Dattels, P., McCaughrin, R., Miyajim, K., Puig, J. (2010), Can you Map Global Financial Stability?.IMF Working Paper, WP/10/145;
[3] Fioramanti, M. (2008), Predicting Sovereign Debt Crises Using Artificial Neural Networks: A Comparative Approach.Journal of Financial Stability,
4 (2), pp 149 – 164;
[4] H. Yin. (2008), The Self-Organizing Maps: Background, Theories, Extensions and Applications. Studies in Computational Intelligence (SCI),
115, pp. 715–762;
[5] Iturriaga, F.J.L., Sanz, I.,P.(2013), Self-organizing Maps as a Tool to Compare Financial Macroeconomic Imbalances: The European, Spanish
and German Case.The Spanish Review of Financial Economics, 11, pp.69-
84;
[6] Kohonen, T. (2013), Essentials of the Self-organizing Map. Neural Networks, 37, pp. 52 – 65;
[7] Miricescu, E. C. (2014), Investigating the Determinants of Long-run Sovereign Rating.Financial Studies,Volume 18, issue 3, pp. 25-32;
[8] Moscalu, M. (2014), The Impact of Interest Rate Spreads for Euro Denominated Loans on the Leverage Ratio of Romanian Listed
Companies. Proceedings of the 16th International Scientific Conference Finance and Risk 2014,vol. 1. Bratislava: Publishing House EKONÓM, pp.
138 – 145;
[9] Oet, M.V., Gramlich, D., Miller, G.L., Ong, S.J. (2010), Early Warning Systems for Systemic Banking Risk: Critical Review and Modeling
Implications.Banks and Bank Systems, Vol. 5 (2);
[10] Reinhart, C.M., Rogoff K.S. (2008), Is the 2007 Sub-prime Financial Crisis so Different? An International Historical Comparison.American
Economic Review. 98 (2), pp. 339 – 344;
[11] Sarlin, P. (2013a), Self-Organizing Time Map: An Abstraction of Temporal Multivariate Patterns.Neurocomputing, 99 (1), pp. 496 – 508;
-
Using Self-organizing Maps for Assessing Systemic Risk. Evidences from the
Global Economic Crisis
_________________________________________________________________
[12] Sarlin, P. (2013b), Decomposing the Global Financial Crisis: A Self-Organizing Time Map.Pattern Recognition Letters,34, pp. 1701 – 1709;
[13] Sarlin, P., Marghescu, D. (2011), Visual Predictions of Currency Crises using Self-Organizing Maps. Intelligent Systems in Accounting, Finance
and Management, 18(1), 15–38;
[14] Sarlin., P., Peltonen, T.(2011), Mapping the State of Financial Stability.ECB Working Papers, No. 1382;
[15] Tsai, C.-F. (2014), Combining Cluster Analysis with Classifier Ensembles to Predict Financial Distress. Information Fusion, 16.